A foundation model for energy and radiation systems built on heterogeneous scientific interfaces
GEODE tests whether a shared scientific model can reuse pretrained computation across incompatible physics interfaces, not just preserve old tasks.
The paper reports one jointly pretrained model spanning cavity flow, radiation dose and elastoplastic stress, then adding heat exchanger and reactor subchannel tasks through a private interface with 2.1% of parameters. Parameter isolation kept earlier predictions unchanged, while unrestricted fine-tuning degraded them by factors of 14-29. The authors say preservation is not proof of reuse: randomized-library controls showed pretrained computation helped only under some task and data conditions. They also warn that full-field relative L2 error can hide spatially meaningful error when field level dominates the norm. ArXiv · AI/CL/LG's note
The paper reports one jointly pretrained model spanning cavity flow, radiation dose and elastoplastic stress, then adding heat exchanger and reactor subchannel tasks through a private interface with 2.1% of parameters. Parameter isolation kept earlier predictions unchanged, while unrestricted fine-tuning degraded them by factors of 14-29. The authors say preservation is not proof of reuse: randomized-library controls showed pretrained computation helped only under some task and data conditions. They also warn that full-field relative L2 error can hide spatially meaningful error when field level dominates the norm. ArXiv · AI/CL/LG's note
score 4